Abstract:Uncertainty estimation is a key factor that makes deep learning reliable in practical applications. Recently proposed evidential neural networks explicitly account for different uncertainties by treating the network's outputs as evidence to parameterize the Dirichlet distribution, and achieve impressive performance in uncertainty estimation. However, for high data uncertainty samples but annotated with the one-hot label, the evidence-learning process for those mislabeled classes is over-penalized and remains hindered. To address this problem, we propose a novel method, Fisher Information-based Evidential Deep Learning ($\mathcal{I}$-EDL). In particular, we introduce Fisher Information Matrix (FIM) to measure the informativeness of evidence carried by each sample, according to which we can dynamically reweight the objective loss terms to make the network more focused on the representation learning of uncertain classes. The generalization ability of our network is further improved by optimizing the PAC-Bayesian bound. As demonstrated empirically, our proposed method consistently outperforms traditional EDL-related algorithms in multiple uncertainty estimation tasks, especially in the more challenging few-shot classification settings.

What problem does this paper attempt to address?

### Problems the Paper Attempts to Solve This paper aims to address key issues in uncertainty estimation within deep learning models. Specifically, existing Evidential Neural Networks (ENNs) can parameterize the Dirichlet distribution by treating network outputs as evidence, thereby explicitly handling different uncertainties and achieving significant performance in uncertainty estimation. However, they have limitations when dealing with samples with high data uncertainty. **The main issues include:** 1. **Over-penalization**: For samples with high data uncertainty but labeled with one-hot encoded labels, the evidence learning process overly penalizes the misclassified categories, hindering the learning of these categories. 2. **Underestimation of Data Uncertainty**: Existing methods tend to underestimate irreducible data uncertainty during training, reducing the usability of uncertainty estimation. 3. **Insufficient Generalization**: Existing methods lack sufficient generalization in challenging tasks, such as few-shot classification settings. To address these issues, the authors propose a new method—Fisher Information-based Evidential Deep Learning (I-EDL). This method introduces the Fisher Information Matrix (FIM) to measure the amount of information each sample carries as evidence and dynamically adjusts the weight of the target loss term, allowing the network to focus more on the representation learning of uncertain categories. Additionally, optimizing the PAC-Bayesian bound further enhances the network's generalization ability. ### Main Contributions 1. **Proposed a New Method**: Combining Fisher Information and Evidential Neural Networks to quantify uncertainty. 2. **Introduced PAC-Bayesian Bound**: Further improving the model's generalization ability. 3. **Outstanding Experimental Performance**: Demonstrated excellent performance in tasks such as confidence evaluation and OOD detection, especially in more challenging few-shot classification settings. ### Key Technologies of the Solution - **Fisher Information Matrix (FIM)**: Used to measure the amount of information each sample carries as evidence and dynamically adjust the weight of the loss term. - **PAC-Bayesian Bound**: Optimizes the model's generalization ability. - **Generative Model**: Assumes the target variable $ y $ follows a multivariate Gaussian distribution and learns model parameters by maximizing the marginal likelihood. With these technologies, the I-EDL method is more robust in handling samples with high data uncertainty and performs excellently in various uncertainty estimation tasks.

Uncertainty Estimation by Fisher Information-based Evidential Deep Learning

Knowledge from Uncertainty in Evidential Deep Learning

Deep Evidential Fusion Network for Image Classification

A Comprehensive Survey on Evidential Deep Learning and Its Applications

Uncertainty Estimation by Density Aware Evidential Deep Learning

Are Uncertainty Quantification Capabilities of Evidential Deep Learning a Mirage?

Is Epistemic Uncertainty Faithfully Represented by Evidential Deep Learning Methods?

TEDL: A Two-stage Evidential Deep Learning Method for Classification Uncertainty Quantification

Revisiting Essential and Nonessential Settings of Evidential Deep Learning

Hyper Evidential Deep Learning to Quantify Composite Classification Uncertainty

Evidential Deep Learning: Enhancing Predictive Uncertainty Estimation for Earth System Science Applications

Uncertainty Regularized Evidential Regression

Region-based evidential deep learning to quantify uncertainty and improve robustness of brain tumor segmentation

Deep Evidential Fusion Network for Medical Image Classification

MEDL-U: Uncertainty-aware 3D Automatic Annotation based on Evidential Deep Learning

Uncertainty-Error correlations in Evidential Deep Learning models for biomedical segmentation

Uncertainty Estimation for 3D Object Detection via Evidential Learning

E-NER: Evidential Deep Learning for Trustworthy Named Entity Recognition

Evidential Uncertainty Quantification: A Variance-Based Perspective

Uncertainty-aware Health Diagnostics Via Class-balanced Evidential Deep Learning

EvCenterNet: Uncertainty Estimation for Object Detection using Evidential Learning